Contributors to the forecasting process (Input):
Users of forecasting (Output):
The Forecasting Lead is responsible for preparing and delivering a range of forecasts that are used by the business to meet planning, regulatory and management obligations. These include minimum and maximum demand forecasts, as well as key inputs including customer growth forecasts and forecasts of distributed energy resources. The Forecasting Lead will act as a key liaison between several AusNet Services teams who rely on, or contribute to the forecasting process.
Electricity demand data are a sequence of time-indexed data, so it is considered as a timeseries. However, this data could have other dimensions. If it is reported for different cross-sections (i.e. geographical or distribution regions), the data could be considered as a panel data. Further, as the cross-sections of demand data are usually spatially interconnected, we have spatial panel data.
First, let’s focus on the timeseries characteristic of our data. Timeseries data could be decomposed to the following components, supposing an additive (linear) model. This decomposition helps us in better understanding our data and forecasting.
where:
To model the above-explined data (timeseries with seasonal effects), Holt-Winters and ARIMA model are used widely; however, there are two points of consideration, in our case. First, Holt-Winters approach does not allow us to directly incorporate exogenous predictors (socio-economic variables) to the model. This is what we could do in the next task of customer growth forecast, but ARIMA does. Second, both these models do best on the timeseries with single seasonal effect. This is where our data shows its complexity, as it potentially is a multiseasonal timeseries. That is multiple seasonal effects (cycles) can be detected, including:
As been indicated, we expect 3 seasonal effects to be detected in our data. Let’s have a look at the below example, using data from Duquesne Light Company.
The intuition from the above figure is that the maximum demand occurs on the “afternoon of Tue/Wed/Thus during the August” of each year; and similarly, the minimum demand is for the “dawn of the Sundays during May” of each year.
Furthermore, as we discussed above, augment our model with exogenous variables could add to our model explanatory power. This could, in summary, include:
Minimum and maximum demand forecasts required for:
Modelling customer growth is somehow different from the previous step. That is although the data could be panel data or spatial panel data, the cross-section dimension of that is the main one, in our forecasting process. Different socio-economic factors would impact customer growth, including:
To model our data, regression models (e.g. multiple regressions, logistic regression) and classification models (e.g. knn) are the candidate ones. Furthermore, as our data are potentially spatially auto-correlated, we may consider spatial autoregression (SAR). Another option would be the geographical weighted regression (GWR), which shows us the spatial heterogeneity in elasticity of our variable (customer growth) to the explanatory variables. The important part here is what geographical segmentation to use. We can use the existing ones (e.g. postcode, LGA, …) or possibly define new ones which represent electricity demand zones the best.
The outputs of this step could be of use in the process of demand forecasting. Hence, it is important to have customer growth forecasting categorised by housing type (i.e. house, townhouse, apartment), commercial area, and industry type. That is each of them adds a different level of energy load to the network.
New customer connections and disconnections (abolishments) to be prepared at the feeder level for:
Similar to the customer growth forecasting, here we have a forecast of cross-sectional data. If we use the lagged (i.e. from previous period) explanatory variables, we can have future (i.e. forecast) values. The main point here is that there are some explanatory variables which changes in them could impose a shock to the system and alter the figures significantly. Hence, we need to regularly monitor them. These include (for photovoltaic (PV) energy):
As a key input to minimum and maximum demand, the Forecasting Lead will be responsible for developing AusNet Services’ view on the growth in distributed energy resources, including:
In each forecasting project, all the processes should be done in a row. But this doesn’t mean that the output (estimated parameters) could not be used several times. Once the best model is selected, we can use its parameters for forecasting future or new values, until we need to (or see it better) to re-do the process.
For the one time forecasts, it could be done using any software or programming language; then the output could be presented as a data file or on a figure (graph, map, …). But for the cases when it needed to be repeated on new out-of-sample data over time, we can design a toolset which has the forecasting parameters in the background and produce output for new data. This could be done in R-Markdown using Shiny Web Apps. R-Markdown is a powerful presentation tool. This, unlike other presentation tools like Tableau which do just presentation, could do calculations and presentation at the same time. It could be even programmed to do all the forecasting process automatically. For instance, it could use the knn regression output and forecast the electricity usage of a new building by comparing its characteristics with the training database. Besides, we can export the forecasting parameters (output) and use them in any other platform (e.g. python, java) used by organisational units.
The next point is the frequency of updating the forecasting model. This depends on the stability of the systems. For instance, in the timeseries forecasting of demands, the trend forecast could be affected by many factors and needs shorter term (i.e. daily) updates, but the seasonal (cyclical) effects may be stable for several months; although testing for potential changes are necessary. Similarly, in case of the causal forecasting of customer growth, as the explnatiry variables are exogenous and could change at any point in time (e.g. have structural breaks), regular updates are necessary. The output of the customer growth forecast then impacts the timeseries forecasting of the above step.
The Forecasting Lead is required to work collaboratively with the Network Innovation Department, which builds the forecasting tools relied upon by the Forecasting Lead. Specifically, the Forecasting Lead is responsible for: - Actively contributing to the development of demand forecasting toolsets - Regularly meeting with key members of the Network Innovation Department to discuss the performance of the tools and suggesting improvements - In conjunction with the Network Planning department, recommending new tools that will result in improved forecasts
The internal stakeholders would have a two-directional relation with the forecasting unit. They, to some extent, provide data used by the forecasting unit. On the other hand, they use the reports and data produced in their planning and policy-making. The point is, in many cases, we need to present outputs less technically. In other words, the presentations (data, graphs) should be in a way (i.g. with the least jargons) that could be understandable by non-technicals (non-statisticians).
Furthermore, the forecasting unit could have meetings with different organisational units, presenting them what the unit does and how could the forecasting help AusNet in obtaining its goals. Then the internal data-related processes could be reviewed in one by one meetings and suggestions and solutions related to forecasting could be provided.
Act as a contact point for a number of teams within AusNet Services who have an interest in the forecasting process or outputs. This includes:
Electricity is generated, used in each region and traded across regions.
High voltage transmission lines transport electricity from generators to electricity distributors, who deliver it to homes and businesses on lower voltage ‘poles and wires’.
The NEM operates on one of the world’s longest interconnected power systems - from Port Douglas in Queensland to Port Lincoln in South Australia - a distance of around 5,000 kilometres.
The NEM generates around 200 terawatt hours of electricity annually, supplying around 80% of Australia’s electricity consumption.
The electricity industry is one of Australia’s largest. It provides energy for industry, businesses and households.
Western Australia and the Northern Territory are not connected to the NEM. They have their own electricity systems and separate regulatory arrangements, although the AEMC also has a role in the Northern Territory.